On Tuesday, April 26, 2011 10:42:17 am Hans Petter Selasky wrote:
> On Tuesday 26 April 2011 16:37:17 John Baldwin wrote:
> > On Tuesday, April 26, 2011 10:27:14 am Warner Losh wrote:
> > > On Apr 26, 2011, at 7:42 AM, John Baldwin wrote:
> > > > - The Giant protection for new-bus should prevent attach/detach from
> > > > running
> > > >
> > > > concurrently I believe (either that or the USB bus itself should
> > > > ensure that the two instances of your device have seperate device_t
> > > > instances with separate softc's, so current attach/detach should not
> > > > matter except that they may both try to talk to the same hardware
> > > > perhaps? In that case that is something the USB bus driver should
> > > > fix by prevent a device from attaching at an existing address until
> > > > any existing device at that address is fully detached).
> > >
> > > I thought that if we held Giant when we're about to go to sleep that we
> > > drop it as a special case. So if any newbus-releated function sleeps,
> > > we can have a situation where attach is running and detach gets called.
> > > There is (or was) some code to cope with this in CardBus, iirc. I'm
> > > surprised there isn't any in USB, since Hans was the one that alerted me
> > > to this issue.
> > Yes, Giant doesn't really provide too much help here. However, the real
> > fix should be in the USB bus, and USB peripheral drivers should not have
> > to worry about handling concurrent attach/detach (they can't really handle
> > it safely anyway).
> All detach/attach/suspend/resume functions on a device tree belonging to the
> same USB controller are executed from a single thread, which is called the
> root HUB thread.
Ok, that should work fine then to serialize the detach and attach.